Brief Communications

Keeping Memory Clear and Stable—The Contribution ofHuman Basal Ganglia and Prefrontal Cortex to WorkingMemory

Bernhard Baier,1 Hans-Otto Karnath,2,3 Marianne Dieterich,4 Frank Birklein,1 Carolin Heinze,1 and Notger G. Muller5,6

1Department of Neurology, University of Mainz, 55131 Mainz, Germany, 2Section of Neuropsychology, Center of Neurology, Hertie-Institute for ClinicalBrain Research, University of Tubingen, 72076 Tubingen, Germany, 3Center for Advanced Brain Imaging, Georgia Institute of Technology, Atlanta, Georgia30332, 4Department of Neurology, Ludwig-Maximilians-University, 81377 Munich, Germany, and 5Department of Neurology and 6German Centre forNeurodegenerative Diseases, Otto-von-Guericke-University, 39120 Magdeburg, Germany

Successful remembering involves both hindering irrelevant information from entering working memory (WM) and actively maintainingrelevant information online. Using a voxelwise lesion– behavior brain mapping approach in stroke patients, we observed that lesions ofthe left basal ganglia render WM susceptible to irrelevant information. Lesions of the right prefrontal cortex on the other hand make itdifficult to keep more than a few items in WM. These findings support basal ganglia–prefrontal cortex models of WM whereby the basalganglia play a gatekeeper role and allow only relevant information to enter prefrontal cortex where this information then is activelymaintained in WM.

IntroductionWhy do some people have a memory like an elephant while thatof others resembles a sieve? One crucial determinant for the in-dividual working memory (WM) capacity is the ability to inhibitirrelevant information from entering memory (Vogel et al.,2005). A recent fMRI study located the competence to filter rel-evant from irrelevant information for memory storage in theprefrontal cortex (PFC) and left basal ganglia (BG) (McNab andKlingberg, 2008). In this and in an earlier study (Todd andMarois, 2004), the right parietal cortex was found to be sensitiveto memory load, i.e., the number of items that have to be remem-bered. Other studies have stressed the role of the PFC in main-taining and manipulating information in WM (D’Esposito et al.,1998; Miller and Cohen, 2001; Muller and Knight, 2006).

It was the aim of the present investigation to clarify whichbrain regions are necessary to sustain filtering and maintenanceof information in WM. While brain activation techniques such asfMRI show which regions are involved in a task, brain disruptiontechniques such as the lesion method enable one to infer that theregion is required (Rorden and Karnath, 2004; Chatterjee, 2005;Muller and Knight, 2006). We thus studied stroke patients withvarying brain lesions due to stroke in a visual spatial workingmemory task applying a voxelwise lesion– behavior brain map-ping (VLBM) analysis. This procedure tests whether the magni-tude of a behavioral variable (i.e., the ability to filter irrelevantinformation and the ability to store information) is significantly

associated with a certain location in the brain (Rorden et al.,2007).

Materials and MethodsThe WM task involved three different conditions: patients had to re-member the position of three red dots without distracting items on dis-play, they had to remember the positions of three red dots while ignoringtwo simultaneously presented yellow dots, or they had to remember theposition of five red dots with no other objects on display. We then com-puted the corrected (hits minus false alarms) hit rate differences betweenthree target trials with versus without distractors to assess filtering abilityand the differences between no distractor trials with five versus threetargets to assess load-sensitive maintenance. Timing of the paradigm canbe depicted from Figure 1.

Sixty-one randomly selected patients with cortical damage due tostroke as demonstrated by magnetic resonance imaging (MRI) were in-vestigated. Thirty-one patients had right-sided (51%) and 30 patientsleft-sided brain damage (49%). Lesions were mainly in the territory of theA. cerebri media, but parts of the territories of the A. cerebri anterior andA. cerebri posterior were also affected (Fig. 2 A, B). Exclusion criteriawere visual field defects, diffuse brain damage, intake of psychoactivedrugs within 24 h before examination, or insufficient communicationabilities. The latter introduced a bias for lesions to be smaller in thelanguage-dominant left hemisphere; the difference, however, was notstatistically different (Mann–Whitney test p � 0.143). Medications thatwere regularly taken were anti-platelet-aggregating drugs, anticoagu-lants, cholesterol synthesis enzyme inhibitors, or anti-hypertensivedrugs; none of these were expected to interfere with cognitive function.Clinical parameters were assessed as described previously (Karnath et al.,2005a). For patients with neglect symptoms (all right-sided lesions), thestimulation monitor was located on the ipsilesional side to reduce theirdirection-specific inattention. With this procedure, our neglect patientsdid not show significantly more misses for targets presented on the leftversus right side (paired t test; p � 0.155). Furthermore, as we analyzeddifferences of accuracy rates rather than absolute values, a deficit in de-

Received March 24, 2010; revised May 12, 2010; accepted June 8, 2010.This work is part of a medical thesis of C.H.Correspondence should be addressed to Dr. Bernhard Baier, Department of Neurology, University of Mainz,

Langenbeckstrasse 1, 55131 Mainz, Germany. E-mail: baierb@uni-mainz.de.DOI:10.1523/JNEUROSCI.1513-10.2010

Copyright © 2010 the authors 0270-6474/10/309788-05$15.00/0

9788 • The Journal of Neuroscience, July 21, 2010 • 30(29):9788 –9792

tecting left-sided stimuli per se would not have a systematic effect. Fi-nally, no lesions usually related to neglect (right temporoparietal corticaljunction) were found to be associated with the processes of interest.Clinical variables are shown in Table 1.

MRI scans were performed on all patients. We used diffusion-weighted imaging (DWI) within the first 48 h after stroke and fluid-attenuated inversion recovery (FLAIR) sequences when imaging wasconducted later than 48 h after the stroke. The DWI sequence comprised38 axial slices with an interslice gap of 3.3 mm. FLAIR images wereacquired with a slice thickness of 1 mm. The mean time between lesionand MRI was 6 d (SD 2.2 d); in other words, the patients were tested in theacute phase before substantial reorganization of brain function wouldmake interpretation of the results more difficult. Moreover, the fact thatrelative instead of absolute behavioral measures were assessed makes itunlikely that unspecific disease effects contributed to the results reportedhere. The boundary of the lesion was delineated directly on the individualMRI image for every single transversal slice using MRIcron software(Rorden et al., 2007) (http://www.sph.sc.edu/comd/rorden/mricron/).Both the scan and the lesion shape were then mapped into stereotaxicspace using the normalization algorithm provided by SPM8 (http://www.fil.ion.ucl.ac.uk/spm/software/spm8). Lesions were presented on a T1-weighted template MRI scan from the Montreal Neurological Institute(MNI). The extension and location of the lesion shapes were controlledby a second experimenter (C.H.) who was blind to the performance ofthe patients in the working memory paradigm. MNI coordinates wereassigned to cerebral structures in general, including subcortical areassuch as the basal ganglia and white matter using an MNI-space utilitysoftware tool (www.ihb.spb.ru/�pet_lab/MSU/MSUMain.html). Onlyvoxels damaged in at least 3% of patients were included into the analysis.

Filtering ability was assessed with a statistical VLBM analysis thattreated the difference in accuracy rates between the no-distraction con-dition with three targets and the distraction condition (i.e., with addi-tional irrelevant items) with the same amount of targets of each patient asthe dependent, continuously measured variable. Accuracy rates were de-fined as corrected hit rates, i.e., hits minus false alarms to account fordifferences in response strategies between subjects. Right- and left-sidedlesions were analyzed separately. Working memory capacity was assessedin a second VLBM analysis whereby the dependent, continuously mea-sured variable was the difference in corrected hit rates between the no-distraction condition with three targets and the no-distraction conditionwith five targets.

We used the nonparametric Brunner Munzel test with the significancelevel set to p � 0.05 (Rorden et al., 2007). To prevent a rise in theprobability of familywise error, we computed a false discovery rate (FDR)correction in all analyses.

ResultsThe patients reached the highest accuracyin the three-target-no-distractor condi-tion (hits: 78%, false alarms: 14%), andlower accuracy in the five-target-no-distractor condition (hits: 67%, falsealarms: 20%) and in the three-target-plus-distractor condition (hits: 65%, falsealarms: 19%). Accordingly, the repeated-measurement ANOVA revealed a maineffect for the factor task (three levels)(F(1,60) � 56.263; p � 0.001). Pairwisecomparisons revealed significant differ-ences between the three-target-nodistractor and both the three-target-plus-distractor and the five-target-no-distractor conditions (paired t test p �0.001).

The VLBM analysis related to filteringability revealed an association of thisfunction with lesions of the left putamen(x � �28, y � �8, z � 13) and adjacent

white matter (x � �28, y � �8, z � 18) (FDR-corrected � levelof p � 0.05) (Fig. 2C). No significant voxels in the right hemi-sphere were identified with this analysis. Figure 2D shows theVLBM analysis for the factor WM capacity, indicating that whitematter lesions of the right frontal lobe (x � 23, y � �2, z � 34) aswell as the inferior frontal gyrus (x � 42, y � 19, z � 13), the headof the caudate nucleus (x � 11, y � 14, z � 9), and the insularcortex (x � 44, y � 11, z � 12; x � 34, y � �8, z � 18) weresignificantly associated with low WM capacity (FDR-corrected �level of p � 0.05). No voxels in the left hemisphere were identi-fied. In an additional analysis, we conducted a VLBM analysis forfactor WM capacity excluding patients with neglect and extinc-tion. Anatomical results remained unaffected (Fig. 2E).

We supplemented the VLBM analysis by using a traditionalapproach where we compared the behavior of patients whoselesions involved the left basal ganglia and right frontal cortex,respectively, with patients that spared these regions. The follow-ing subgroups were built for this analysis: 10 left-sided patientswith lesions of predominantly the left putamen, nine right-sidedpatients with predominantly prefrontal lesions, 20 left-sided pa-tients without predominantly basal ganglia damage, and 22 right-sided patients without prefrontal damage. However, it has to benoted that such a traditional approach of building post hoc sub-groups for lesion analysis is very problematic (cf. Rorden andKarnath, 2004). In the present case, the post hoc subgroups showan obvious confound with lesion size. Patients with right hemi-sphere lesions with frontal involvement had larger lesions (lesionvolume 57.8 cc) than patients without frontal involvement (le-sion volume 8.35 cc; independent samples t test t � 5.12; p �0.001). Similarly, patients with left hemisphere lesions with pu-tamen involvement (lesion volume 10.3 cc) showed larger lesionsthan patients without affection of the putamen (lesion volume1.97 cc; independent samples t test t � 2.52; p � 0.018). From aconceptual standpoint, given that patients with large right-hemisphere lesions are roughly more likely to have frontal in-volvement, then the threshold criterion will select more subjectswith frontal involvement, and thus bias the results to showfrontal effects. Again, along these same lines, patients withlarge left-hemisphere lesions are more likely to have putameninvolvement, which would again bias the statistics. This caveat

Figure 1. Experimental design: Trials started with the presentation of the 12 position grid and either three red dots, five reddots, or three red dots with two yellow dots. Subjects had to remember the positions of the red dots only.

Baier et al. • Working Memory and Basal Ganglia J. Neurosci., July 21, 2010 • 30(29):9788 –9792 • 9789

needs to be kept in mind when we assesshow the four subgroups’ behavior was af-fected by adding distracters to the to-be-memorized items (filtering deficit) and byincreasing memory load (memory defi-cit), respectively (Fig. 3A,B). The mixedANOVA analysis revealed an interaction be-tween group (four levels) and cognitive def-icit (filtering vs memory) with F � 19.729;p � 0.001. As could be expected from Figure3, a post hoc analysis (Bonferroni cor-rected) indicated that the 10 patientswith lesions affecting the left-sided pu-tamen showed worse filtering abilitythan the left-sided lesion patients with-out involvement of the putamen ( p �0.001). Furthermore, the nine patientswith lesions involving the right-sided pre-frontal structures were stronger affectedby increased memory load than the otherright-sided patients ( p � 0.001).

DiscussionWe observed that lesions of the left puta-men and surrounding white matter spe-cifically impaired performance when theto-be-remembered targets were presentedtogether with task-irrelevant items. Le-sions of the right frontal lobe on the otherhand impaired performance especiallywhen memory load was increased fromthree to five items. Thus, we conclude thatthe left putamen is involved in hinderingirrelevant information from enteringworking memory, whereas the right pre-frontal cortex is crucial for active main-taining relevant information online inWM. These findings substantiate compu-tational models of WM (Frank et al., 2001;Gruber et al., 2006; Hazy et al., 2006;O’Reilly and Frank, 2006). These modelsproposed that the striatum should play adopamine-dependent gate-keeping func-tion and should control the informationflow into working memory whereas mem-ory representations are maintained inPFC. In detail, one model suggested thatparallel loops interconnect the frontalcortex with the basal ganglia.

Due to direct signals originating from“Go” neurons in the dorsal striatum, adisinhibition of the frontal cortex resultsvia an inhibition of the substantia nigrapars reticulata leading to a gating-likemodulation and updating of WM repre-sentations in PFC. An indirect pathwayinvolving “No Go” neurons in the dorsalstriatum that inhibit the inhibitory globuspallidus supported by neurons in the sub-thalamic nucleus that excite the substantianigra pars reticulata acts as a counter-player. By these mechanisms, the gate toWM in PFC is locked for irrelevant infor-

Figure 2. A, Overlay lesion plot of all 30 left-sided lesion patients. The number of overlapping lesions is illustrated by differentcolors coding increasing frequencies from violet (n � 1) to red (n � 30). B, Overlay lesion plot of all 31 right-sided lesion patients.The number of overlapping lesions is illustrated by different colors coding increasing frequencies from violet (n � 1) to red (n �31). C, VLBM analysis of filtering ability, left-sided brain lesions. The behavioral variable was determined by calculating thecorrected hit rate differences between the conditions with versus without distraction and three targets. Lesions of the putamenwere associated with deficits in filtering ability (FDR-corrected � level of p � 0.05). D, VLBM analysis of WM capacity, right-sidedbrain lesions. The behavioral variable was determined as the corrected hit rate difference between the no-distraction conditionswith five versus three targets. Right-sided lesions of the inferior frontal gyrus and insular cortex were associated with poor WMcapacity (FDR-corrected � level of p � 0.05). Talairach z-coordinates (Talairach and Tournoux, 1988) of each transverse section aregiven. E, VLBM analysis of WM capacity, right-sided brain lesion patients without neglect.

9790 • J. Neurosci., July 21, 2010 • 30(29):9788 –9792 Baier et al. • Working Memory and Basal Ganglia

mation (Frank et al., 2001; Hazy et al., 2006). Transferring thismodel to our data, lesion of the left putamen would cause disrup-tion of the “Go” and “No Go” neurons of the striatum leading toan imbalance of the gating mechanism. Alternatively or addition-ally, the striatal lesions may have caused a dysfunction of distantcortical areas via diaschisis (Karnath et al., 2005b). Structural

MRI scans might not necessarily show the full functional extentof a lesion. (Cortical) areas that appear structurally intact in an-atomical scans may not necessarily be functioning normally dueto an abnormal perfusion. Therefore, normalized perfusion-weighted imaging (Karnath et al., 2005b), which measures theamount and latency of blood flow in certain regions, provides apromising tool to address these issues in future studies.

Likewise, as the prefrontal lesions mainly affected the subgyralwhite matter, malfunctioning in distant areas might have con-tributed to the memory disturbance in these patients as well. Thefinding that the relevant lesions were observed in inferior andinsular regions rather than in more dorsal regions may partly beexplained by the fact that these lesions were more common in oursample. However, other data indeed indicate that more inferiorfrontal regions support memory maintenance, whereas dorsalfrontal areas are found to be involved in manipulation of infor-mation within WM, a process that was not required in our study(D’Esposito et al., 2000). Finally, with regard to the insular cortexand the caudate nucleus, it was shown in animal studies that thisarea is part of an orbital prefrontal network that might also play arole in working memory processes (Levy et al., 1997; Saleem et al.,2008; White, 2009).

Unpredicted by the models, we observed a hemispheric spe-cialization, which at least for the BG is in line with a recent fMRIstudy (McNab and Klingberg, 2008). In this study a significantcorrelation between individual WM capacity and filtering set ac-tivity (operationalized by the signal difference between distrac-tion vs no-distraction trials) was observed in left BG.

Discrepant to our study, two fMRI studies identified the pari-etal, not the frontal cortex, to be memory load sensitive (Toddand Marois, 2004; McNab and Klingberg, 2008). One explana-tion might be that the parietal cortex is sensitive to overall cogni-tive load, but that this is not specific to memory load per se.Increasing the number of items in a display, for example, alsoincreases attentional load. Indeed, an fMRI study by Tomasi et al.(2007) found that parietal cortex was driven both by cognitiveload in an attention and memory task, whereas PFC was load-driven only in the WM task. This may indicate that PFC is morespecifically related to WM—an assumption also supported by theexisting computational models of WM (Gruber et al., 2006; Hazyet al., 2006; O’Reilly and Frank, 2006). This notion does not ruleout the involvement of PFC in attention per se. For example, thePFC might have been involved in attentional processes that wererelevant in both the distracter absent and present trials. In thatcase the PFC’s role would have been concealed in our analysis thatwas based on differences between these conditions.

A possible limitation might be that performance in acutestroke patients is vulnerable to other factors such as, e.g., concen-tration. However, the patients in the present study were (1) thor-oughly selected to include only vigilant patients in the study and(2) assessed based on differences in accuracy rates so that generaldeficits would have cancelled each other out. Furthermore, test-ing patients in the chronic phase of their disease would haveintroduced another, possibly more severe, confound: reorganiza-tion of brain function and structure. Thus, testing patients in theacute phase of disease may allow more valid conclusions regard-ing the cortical function in the normal brain.

ReferencesChatterjee A (2005) A madness to the methods in cognitive neuroscience?

J Cogn Neurosci 17:847– 849.D’Esposito M, Aguirre GK, Zarahn E, Ballard D, Shin RK, Lease J (1998)

Functional MRI studies of spatial and nonspatial working memory. BrainRes Cogn Brain Res 7:1–13.

Table 1. Demographic and clinical data of all left- and right-brain-damagedpatients

Right-sided lesions Left-sided lesions

Number 31 30Sex 17 f; 14 m 10 f; 20 mEtiology 31 infarcts 29 infarcts, 1 bleedingAge (years) �median (range)� 68 (20 – 84) 68 (32– 85)Time interval lesion-clinical examination

(d) �median (range)�7 (2–15) 6 (2–24)

Lesion volume (in cc) �median (range)� 10.7 (0.1– 85.4) 2.9 (0.25–139.6)Contralesional paresis (MRC scale)

�median (range)�4 (0 –5) 5 (1–5)

Visual extinction (% present) 10 0Neglect (% present) 13 0MMSE �median (range)� 25 (18 –30) 27 (18 –30)

f, Female; m, male; MMSE, Mini-Mental State Examination.

Figure 3. Subgroup analysis comparing patients with left-sided lesions involving the puta-men (n � 10) (black bar), left-sided lesion patients without affection of the putamen (n � 20)(white bar), right-sided patients with lesions encompassing frontal structures such as the infe-rior frontal gyrus (n � 9) (bright gray bar), and right-sided lesion patients without affection ofthe frontal lobes (n � 22) (dark gray bar). A, Filtering deficit. The y-axis represents the differ-ence in accuracy rates (corrected hit rate) between the condition with three memory items andno distracters and the condition with three memory items and two distractors. A large valueindicates that subjects was strongly impaired by the presence of distractors, i.e., has low filter-ing ability. Error bars indicate the SEM. B, Memory deficit. The y-axis represents the difference inaccuracy between the condition with three memory items and the condition with five items(both without distractors). A large value therefore indicates that subjects had difficulties whenmemory load was increased. Error bars indicate the SEM.

Baier et al. • Working Memory and Basal Ganglia J. Neurosci., July 21, 2010 • 30(29):9788 –9792 • 9791

D’Esposito M, Postle BR, Rypma B (2000) Prefrontal cortical contributions toworking memory: evidence from event-related fMRI studies. Exp BrainRes 133:3–11.

Frank MJ, Loughry B, O’Reilly RC (2001) Interactions between frontal cor-tex and basal ganglia in working memory: a computational model. CognAffect Behav Neurosci 1:137–160.

Gruber AJ, Dayan P, Gutkin BS, Solla SA (2006) Dopamine modulation inthe basal ganglia locks the gate to working memory. J Comput Neurosci20:153–166.

Hazy TE, Frank MJ, O’Reilly RC (2006) Banishing the homunculus: makingworking memory work. Neuroscience 139:105–118.

Karnath HO, Baier B, Nagele T (2005a) Awareness of the functioning ofone’s own limbs mediated by the insular cortex. J Neurosci 25:7134 –7138.

Karnath HO, Zopf R, Johannsen L, Fruhmann Berger M, Nagele T, Klose U(2005b) Normalised perfusion MRI to identify common areas of dys-function: patients with basal ganglia neglect. Brain 128:2462–2469.

Levy R, Friedman HR, Davachi L, Goldman-Rakic PS (1997) Differentialactivation of the caudate nucleus in primates performing spatial and non-spatial working memory tasks. J Neurosci 17:3870 –3882.

McNab F, Klingberg T (2008) Prefrontal cortex and basal ganglia controlaccess to working memory. Nat Neurosci 11:103–107.

Miller EK, Cohen JD (2001) An integrative theory of prefrontal cortex func-tion. Annu Rev Neurosci 24:167–202.

Muller NG, Knight RT (2006) The functional neuroanatomy of working mem-ory: contributions of human brain lesion studies. Neuroscience 139:51–58.

O’Reilly RC, Frank MJ (2006) Making working memory work: a computa-tional model of learning in the prefrontal cortex and basal ganglia. NeuralComput 18:283–328.

Rorden C, Karnath HO (2004) Using human brain lesions to infer function:a relic from a past era in the fMRI age? Nat Rev Neurosci 5:813– 819.

Rorden C, Karnath HO, Bonilha L (2007) Improving lesion-symptom map-ping. J Cogn Neurosci 19:1081–1088.

Saleem KS, Kondo H, Price JL (2008) Complementary circuits connect-ing the orbital and medial prefrontal networks with the temporal,insular, and opercular cortex in the macaque monkey. J Comp Neurol506:659 – 693.

Talairach J, Tournoux P (1988) Co-planar stereotaxic atlas of the humanbrain: 3-dimensional proportional system—an approach to cerebral im-aging. New York: Thieme.

Todd JJ, Marois R (2004) Capacity limit of visual short-term memory inhuman posterior parietal cortex. Nature 428:751–754.

Tomasi D, Chang L, Caparelli EC, Ernst T (2007) Different activation pat-terns for working memory load and visual attention load. Brain Res1132:158 –165.

Vogel EK, McCollough AW, Machizawa MG (2005) Neural measures revealindividual differences in controlling access to working memory. Nature438:500 –503.

White NM (2009) Some highlights of research on the effects of caudatenucleus lesions over the past 200 years. Behav Brain Res 199:3–23.

9792 • J. Neurosci., July 21, 2010 • 30(29):9788 –9792 Baier et al. • Working Memory and Basal Ganglia